Skip to content

ext/pcntl (ZTS): a process-directed signal delivered to a non-PHP thread crashes in pcntl_signal_handler #25541

Open

Description

@phcorp

Description

On a ZTS build configured with --disable-zend-signals (as Homebrew's php-zts is), pcntl_signal_handler() is installed directly with sigaction() and reads PCNTL_G(...) without checking that it is running on a PHP thread. (With zend signals enabled, zend_signal_handler_defer() already checks tsrm_is_managed_thread() before anything reaches pcntl.) A process-directed signal (kill(), alarm()) may be delivered to any thread that does not block it. On macOS, every PHP process has libdispatch worker threads. When the signal lands on one of those threads, the handler dereferences that thread's (absent) TSRM globals.

<?php
// ZTS only: SIGALRM delivered to a non-PHP thread crashes in pcntl_signal_handler.
pcntl_async_signals(true);
pcntl_signal(SIGALRM, static function (): void {});
function weigh(string $s): int { return strlen($s) % 7; }
$pid = getmypid();
$killer = proc_open(['/bin/sh', '-c', "while kill -ALRM $pid 2>/dev/null; do :; done"], [], $pipes);
$end = hrtime(true) + 20_000_000_000;
while (hrtime(true) < $end) {
    $rows = [];
    for ($i = 0; $i < 200; $i++) { $rows[] = str_repeat('x', $i % 13); }
    usort($rows, static fn (string $a, string $b): int => weigh($a) <=> weigh($b));
}
pcntl_signal(SIGALRM, SIG_IGN);
proc_terminate($killer, 9);
echo "survived\n";

Resulted in this output (exit 139, intermittent: about 1 to 2 runs in 6 under a 20 s storm):

* thread #2, stop reason = EXC_BAD_ACCESS (code=1, address=0x0)
    frame #0: php`pcntl_signal_handler + 32
    frame #1: libsystem_platform.dylib`_sigtramp + 56
  thread #1: php`zend_array_destroy + 164, queue = 'com.apple.main-thread'

But I expected this output instead:

survived

In a larger application, the same race shows up as zend_mm_heap corrupted (exit 134) or as a hash lookup that never ends (zend_hash_find spinning). Both appear when a single pcntl_alarm() handler fires while the main thread is busy.

The NTS build of the same version never crashes (0 of 6 runs; also 0 of 6 on an NTS debug build), which is consistent with NTS globals being process-wide.

A storm makes it likely because pcntl_signal_dispatch() blocks every signal on the main thread while it runs handlers, so the kernel has to pick another thread for a process-directed signal that arrives in that window.

A fix along the lines of zend_signal_handler_defer(): under ZTS, when !tsrm_is_managed_thread(), don't touch PCNTL_G(), and forward the signal with pthread_kill() to the thread that installed the handler. I have a PR for this, with a deterministic test that raises the signal on a raw pthread through a new zend_test helper.

PHP Version

PHP 8.5.12 (cli) (ZTS), Homebrew php-zts (NTS 8.5.11 for comparison)

Operating System

macOS (arm64)

Activity

  1. phcorp commented on Nov 18, 2026

    @phcorp
    Author

    Update after working on GH-25542. Some of the description above needs correcting:

    Two related problems are tracked separately: GH-25548 and GH-25549.

  2. phcorp commented on Nov 18, 2026

    @phcorp
    Author

    GH-25542 now fixes this issue together with GH-25548 and GH-25549, which came out of the same investigation. Its description documents the signal policy and the remaining limitations.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions